Data-efficient control policy search using residual dynamics learning

In this work, we propose a model-based and data efficient approach for reinforcement learning. The main idea of our algorithm is to combine simulated and real rollouts to efficiently find an optimal control policy. While performing rollouts on the robot, we exploit sensory data to learn a probabilistic model of the residual difference between the measured state and the state predicted by a simplified model. The simplified model can be any dynamical system, from a very accurate system to a simple, linear one. The residual difference is learned with Gaussian processes. Hence, we assume that the difference between real and simplified model is Gaussian distributed, which is less strict than assuming that the real system is Gaussian distributed. The combination of the partial model and the learned residuals is exploited to predict the real system behavior and to search for an optimal policy. Simulations and experiments show that our approach significantly reduces the number of rollouts needed to find an optimal control policy for the real system.